Information processing device, information processing program, and information processing method

Abstract

An information processor can logically support prediction based on past statistical information even though the information contains qualitative or non-numerical data. The processor determines whether an input pattern corresponding to an input object (a determination target) belongs to a specific class among multiple classes, based on feature subsets of any combination of a plurality of features, each feature comprises multiple categories. The processor includes a storage storing the input pattern corresponding to the input object and samples corresponding to respective sample objects and a classification determiner determining whether the input pattern belongs to the specific class. The classification determiner calculates a first conditional probability and a second conditional probability based on the number of the samples belonging to each category of the respective features, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective feature for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features for a non-specific class which is a class other than the specific class among classes, and the number of the samples is counted for each class based on the feature information on the samples and the class label information on the samples, and determines whether the input pattern belongs to the specific class based on the feature information on the input pattern, the first conditional probability and the second conditional probability.

The invention claimed is: 
     
       1. An information processor determining whether an input pattern corresponding to an input object belongs to a specific class among multiple classes, based on feature subsets of any combination of a plurality of features, each feature comprises multiple categories, the information processor comprising:
 a storage storing the input pattern corresponding to the input object and samples corresponding to respective sample objects; and 
 a classification determiner determining whether the input pattern belongs to the specific class based on the categories of the respective features into which data of the input pattern are classified, wherein the input pattern is associated with feature information indicating the categories of the respective features into which the data of the input pattern are classified, 
 each of samples is associated with feature information indicating the categories of the respective features into which the data of the samples are classified and class label information indicating whether the samples belong to the specific class, and 
 the classification determiner 
 calculates a first conditional probability and a second conditional probability based on the number of the samples belonging to each category of the respective features, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features for a non-specific class which is a class other than the specific class among classes, and the number of the samples is counted for each class based on the feature information on the samples and the class label information on the samples, and 
 determines whether the input pattern belongs to the specific class based on the feature information on the input pattern, the first conditional probability and the second conditional probability. 
 
     
     
       2. The information processor according to  claim 1 , wherein
 the classification determiner 
 calculates a first posterior probability that the input pattern belongs to the specific class based on the feature information on the input pattern, a first prior probability of occurrence of the specific class and the first conditional probability, 
 calculates a second posterior probability that the input pattern belongs to the non-specific class based on the feature information on the input pattern, a second prior probability of occurrence of the non-specific class and the second conditional probability, and 
 determines whether the input pattern belongs to the specific class by comparing the first posterior probability with the second posterior probability. 
 
     
     
       3. The information processor according to  claim 1 , wherein qualitative features are comprised in the features. 
     
     
       4. The information processor according to  claim 3 , wherein quantitative features are comprised in the features. 
     
     
       5. The information processor according to  claim 1 , further comprising:
 a feature selector selecting a discriminant feature subset from the feature subsets, 
 wherein the classification determiner determines whether the input pattern belongs to the specific class based on the categories of the respective discriminant features included in the discriminant feature subset into which the data of the input pattern are classified. 
 
     
     
       6. The information processor according to  claim 5 , wherein the feature selector selects the discriminant feature subset based on the first conditional probability and the second conditional probability. 
     
     
       7. The information processor according to  claim 6 , wherein,
 the feature selector comprises:
 (a) a sample-data extractor extracting part of the samples as training samples and the remaining samples as test samples, from samples; 
 (b) a classification estimator estimating whether the test samples corresponding to the test objects belong to the specific class, based on the feature information on the training samples, the class label information on the training samples, and the feature information on the test samples, for respective feature subsets of any of the features, and determining the correctness of the estimation based on the class label information on the test samples; and 
 (c) a discriminant feature determiner determining the discriminant feature subset based on the correctness of the estimation for each feature subset, and 
 
 the classification estimator estimates whether the test samples corresponding to the test objects belong to the specific class based on the first conditional probability and the second conditional probability. 
 
     
     
       8. The information processor according to  claim 7 , wherein
 the sample-data extractor extracts the test samples from the samples, 
 the classification estimator outputs judgement information indicating the correctness of the estimation of the respective test samples based on the first conditional probability and the second conditional probability, and 
 the discriminant feature determiner determines the discriminant feature subset based on the judgement information on the respective test samples. 
 
     
     
       9. A non-transitory computer readable medium encoded with a program for processing information instructing a computer to function as the information processor according to  claim 1 . 
     
     
       10. A method of processing information carried out by an information processor determining whether an input object belongs to a specific class among multiple classes, based on feature subsets of any combination of a plurality of features, each feature comprises multiple categories, the information processor comprising:
 a storage storing input pattern corresponding to the input object and samples corresponding to respective sample objects; and 
 a classification determiner determining whether the input pattern belongs to the specific class based on the categories of the respective features into which data of the input pattern are classified, the input pattern is associated with feature information indicating the categories of the respective features into which the data of the input pattern are classified, each of samples is associated with feature information indicating the categories of the respective features into which the data of the samples are classified and class label information indicating whether the samples belong to the specific class, the method carried out by the information processor comprising: 
 a step of calculating a first conditional probability and a second conditional probability based on the number of the samples belonging to each category of the respective features, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features for a non-specific class which is a class other than the specific class among classes, and the number of the samples is counted for each class based on the feature information on the samples and the class label information on the samples; and 
 a step of determining whether the input pattern belongs to the specific class based on the feature information on the input pattern, the first conditional probability and the second conditional probability. 
 
     
     
       11. An information processor determining whether an input pattern corresponding to an input object belongs to a specific class among multiple classes, based on discriminant features selected from multiple features with multiple categories, the information processor comprising:
 a feature selector selecting the discriminant features from features; 
 a classification determiner determining whether the input pattern corresponding to the input object belongs to the specific class, based on the categories of the respective features included in the discriminant features into which the data of the input pattern are classified; and 
 a storage storing feature information indicating the categories of the respective features into which the data of samples are classified and class label information indicating whether the samples corresponding to sample objects belong to the specific class, the feature information and the class label information being provided for the samples corresponding to the respective sample objects used for selection of the discriminant features, 
 the feature selector comprising:
 (a) a sample-data extractor extracting part of the samples as training samples and the remaining samples as test samples, from the samples; 
 (b) a classification estimator estimating whether the test samples corresponding to test objects belong to the specific class based on the feature information on the training samples, the class label information on the training samples, and the feature information on the test samples, for respective feature subsets of any combination of the features, and determining the correctness of the estimation based on the class label information on the test samples; and 
 (c) a discriminant feature determiner determining the discriminant features based on the correctness of the estimation for each of feature subsets, wherein, 
 
 the classification estimator 
 (b-1) calculates a first conditional probability and a second conditional probability based on the number of the training samples belonging to each category of the respective features, the first conditional probability is a probability that data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for a non-specific class which is a class other than the specific class among classes, and the number of the training samples is counted for each class based on the feature information on the training samples and the class label information on the training samples, 
 (b-2) calculates a first posterior probability that the test sample corresponding to the test object belongs to the specific class based on the feature information on the test sample, a first prior probability of occurrence of the specific class and the first conditional probability, 
 (b-3) calculates a second posterior probability that the test sample corresponding to the test object belongs to the non-specific class based on the feature information on the test sample, a second prior probability of occurrence of the non-specific class and the second conditional probability, 
 (b-4) outputs classification information indicating a result of an estimation on whether the test samples corresponding to the test objects belong to the specific class by comparing the first posterior probability with the second posterior probability, and 
 (b-5) outputs judgement information indicating the correctness of the estimation by comparing the classification information of the test samples with the class label information on the test samples, and 
 the discriminant feature determiner 
 (c-1) specifies the feature subsets corresponding to the judgement information satisfying a predetermined condition among the judgement information of the respective feature subsets, and 
 (c-2) determines the features included in the specified feature subsets as the discriminant features. 
 
     
     
       12. The information processor according to  claim 11 , wherein the classification estimator calculates the first prior probability and the second prior probability. 
     
     
       13. The information processor according to  claim 11 , wherein,
 the sample-data extractor extracts the test samples from the samples, 
 the classification estimator outputs the judgement information for each of the test samples, and 
 the discriminant feature determiner determines the discriminant features, based on the judgement information of each of the test samples. 
 
     
     
       14. The information processor according to  claim 13 , wherein, the classification estimator
 calculates the first conditional probability, the second conditional probability, the first posterior probability and the second posterior probability and outputs the classification information, for each of the test samples, and 
 outputs the judgement information for each of the test samples, based on the classification information of each of the test samples. 
 
     
     
       15. The information processor according to  claim 11 , wherein the first prior probability is equal to the second prior probability. 
     
     
       16. A non-transitory computer readable medium encoded with a program for processing information instructing a computer to function as the information processor according to  claim 11 . 
     
     
       17. A method of processing information carried out by an information processor determining whether an input pattern corresponding to an input object belongs to a specific class among multiple classes, based on discriminant features selected from multiple features with a plurality of categories, the information processor comprising:
 a feature selector selecting the discriminant features from the multiple features; 
 a classification determiner determining whether the input pattern corresponding to the input object belongs to the specific class based on the categories of the respective features into which data of the input pattern are classified; and 
 a storage storing feature information indicating the categories of the respective features into which the data of samples are classified and class label information indicating whether the samples corresponding to sample objects belong to the specific class, the feature information and the class label information being provided for the samples corresponding to the respective sample objects used for selection of the discriminant features, the method comprising: 
 (a) a sample-data extraction step of extracting part of the samples as training samples and the remaining samples as test samples, from the samples; 
 (b) a class estimation step of estimating whether the test samples corresponding to test objects belong to the specific class based on the feature information on the training samples, the class label information on the training samples, and the feature information on the test samples, for respective feature subsets of any combination of the features, and determining correctness of the estimation based on the class label information on the test samples; and 
 (c) a discriminant feature determination step of determining the discriminant features based on the determined result of the correctness of the estimation for each of feature subsets, wherein, 
 the class estimation step comprising:
 (b-1) a step of calculating a first conditional probability and a second conditional probability based on the number of the training samples belonging to each category of the respective features, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for a non-specific class which is a class other than the specific class among classes, and the number of the training samples is counted for each class based on the feature information on the training samples and the class label information on the training samples, 
 (b-2) a step of calculating the first posterior probability that the test sample corresponding to the test object belongs to the specific class based on the feature information on the test sample, the first prior probability of occurrence of the specific class and the first conditional probability, 
 (b-3) a step of calculating a second posterior probability that the test sample corresponding to the test object belongs to the non-specific class based on the feature information on the test sample, a second prior probability of occurrence of the non-specific class and the second conditional probability, 
 (b-4) a step of outputting classification information indicating a result of an estimation on whether the test samples corresponding to the test objects belong to the specific class by comparing the first posterior probability with the second posterior probability, and 
 (b-5) a step of outputting judgement information indicating the correctness of the estimation by comparing the classification information on the test samples with the class label information on the test samples, and the discriminant feature determination steps comprising: 
 (c-1) a step of specifying the feature subsets corresponding to the judgement information satisfying a predetermined condition among the judgement information of the respective feature subsets, and 
 (c-2) a step of determining the features included in the specified feature subsets, as the discriminant features. 
 
 
     
     
       18. An information processor selecting a discriminant feature subset used for determining whether an input object belongs to a specific class among multiple classes, the information processor comprising:
 a feature selector selecting the discriminant feature subset from multiple feature subsets of any combination of a plurality of features, each feature comprises multiple categories; and 
 a storage storing feature information indicating the categories of the respective features into which the data of samples are classified and class label information indicating whether the samples corresponding to sample objects belong to the specific class, the feature information and the class label information being provided for the samples corresponding to the respective sample objects used for selection of the discriminant feature subset, 
 the feature selector comprising:
 (a) a sample-data extractor extracting part of the samples as training samples and the remaining samples as test samples, from the samples; 
 (b) a classification estimator estimating whether the test samples corresponding to test objects belong to the specific class based on the feature information on the training samples, the class label information on the training samples, and the feature information on the test samples, for respective feature subsets, and determining the correctness of the estimation based on the class label information on the test samples; and 
 (c) a discriminant feature determiner determining the discriminant feature subset based on the correctness of the estimation for each of feature subsets, wherein, 
 
 the classification estimator 
 (b-1) calculates a first conditional probability and a second conditional probability based on the number of the training samples belonging to each category of the respective features, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for a non-specific class which is a class other than the specific class among classes, and the number of the training samples is counted for each class based on the feature information on the training samples and the class label information on the training samples, 
 (b-2) calculates a first posterior probability that the test sample corresponding to the test object belongs to the specific class based on the feature information on the test sample, a first prior probability of occurrence of the specific class and the first conditional probability, 
 (b-3) calculates a second posterior probability that the test sample corresponding to the test object belongs to the non-specific class based on the feature information on the test sample, a second prior probability of occurrence of the non-specific class and the second conditional probability, 
 (b-4) outputs classification information indicating a result of an estimation on whether the test samples corresponding to the test objects belong to the specific class by comparing the first posterior probability with the second posterior probability, and 
 (b-5) outputs judgement information indicating the correctness of the estimation by comparing the classification information of the test samples with the class label information on the test samples, and the discriminant feature determiner determines the feature subset as the discriminant feature subset, the feature subset corresponds to the judgement information satisfying a predetermined condition among the judgement information of the respective feature subsets. 
 
     
     
       19. The information processor according to  claim 18 , wherein the classification estimator calculates the first prior probability and the second prior probability. 
     
     
       20. The information processor according to  claim 18 , wherein the first prior probability is equal to the second prior probability. 
     
     
       21. The information processor according to  claim 18 , wherein,
 the sample-data extractor extracts the test samples from the samples, 
 the classification estimator outputs the judgement information for each of the test samples, and 
 the discriminant feature determiner determines the discriminant feature subsets based on the judgement information of each of the test samples. 
 
     
     
       22. The information processor according to  claim 18 , wherein, the classification estimator
 calculates the first conditional probability, the second conditional probability, the first posterior probability and the second posterior probability and outputs the classification information, for each of the test samples, and 
 outputs the judgement information for each of the test samples, based on the classification information of each of the test samples. 
 
     
     
       23. A non-transitory computer readable medium encoded with a program for processing information instructing a computer to function as the information processor according to  claim 18 . 
     
     
       24. A method of processing information carried out by an information processor selecting a discriminant feature subset used for determining whether an input object belongs to a specific class among multiple classes, the information processor comprising:
 a feature selector selecting the discriminant feature subset from multiple feature subsets of any combination of a plurality of features, each feature comprises multiple categories; and 
 a storage storing feature information indicating the categories of the respective features into which the data of samples are classified and class label information indicating whether the samples corresponding to sample objects belong to the specific class, the feature information and the class label information being provided for the samples corresponding to the respective sample objects used for selection of the discriminant feature subset, the method comprising: 
 (a) a sample-data extraction step of extracting part of the samples as training samples and the remaining samples as test samples, from the samples; 
 (b) a classification estimation step of estimating whether the test samples corresponding to test objects belong to the specific class based on the feature information on the training samples, the class label information on the training samples, and the feature information on the test samples, for respective feature subsets, and determining the correctness of the estimation based on the class label information on the test samples; and 
 (c) a discriminant feature determination step of determining the discriminant feature subset based on the correctness of the estimation for each of feature subsets, wherein, 
 the classification estimation step comprising:
 (b-1) a step of calculating a first conditional probability and a second conditional probability based on the number of the training samples belonging to each category of the respective features, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for a non-specific class which is a class other than the specific class among classes, and the number of the training samples is counted for each class based on the feature information on the training samples and the class label information on the training samples, 
 (b-2) a step of calculating a first posterior probability that the test sample corresponding to the test object belongs to the specific class based on the feature information on the test sample, a first prior probability of occurrence of the specific class and the first conditional probability, 
 (b-3) a step of calculating a second posterior probability that the test sample corresponding to the test object belongs to the non-specific class based on the feature information on the test sample, a second prior probability of occurrence of the non-specific class and the second conditional probability, 
 (b-4) a step of outputting classification information indicating a result of an estimation on whether the test samples corresponding to the test objects belong to the specific class by comparing the first posterior probability with the second posterior probability, and 
 (b-5) a step of outputting judgement information indicating the correctness of the estimation by comparing the classification information of the test samples with the class label information on the test samples, and 
 
 the discriminant feature determination step comprising: 
 a step of determining the feature subset as the discriminant feature subset, the feature subset corresponds to the judgement information satisfying a predetermined condition among the judgement information of the respective feature subsets. 
 
     
     
       25. An information processor determining whether an input object belongs to a specific class among multiple classes, based on discriminant feature subset consisted of a plurality of discriminant features, each discriminant feature comprises multiple categories, the information processor comprising:
 a storage storing an input pattern corresponding to the input object, a first conditional probability and a second conditional probability, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective discriminant features for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective discriminant features for a non-specific class which is a class other than the specific class among classes; and 
 a classification determiner determining whether the input object belongs to the specific class based on the categories of the respective discriminant features into which data of the input pattern are classified, wherein 
 the input pattern is associated with feature information indicating the categories of the respective discriminant features into which the data of the input pattern are classified, and 
 the classification determiner 
 calculates a first posterior probability that the input object belongs to the specific class based on the feature information on the input pattern, a first prior probability of occurrence of the specific class and the first conditional probability, 
 calculates a second posterior probability that the input object belongs to the non-specific class based on the feature information on the input pattern, a second prior probability of occurrence of the non-specific class and the second conditional probability, and 
 determines whether the input object belongs to the specific class by comparing the first posterior probability with the second posterior probability. 
 
     
     
       26. The information processor according to  claim 25 , wherein the discriminant feature subset is selected from feature subsets of any combination of a plurality of features by a same or different information processor configured for selecting a discriminant feature subset used for determining whether an input object belongs to a specific class among multiple classes, the same or different information processor comprising:
 a feature selector selecting the discriminant feature subset from multiple feature subsets of any combination of a plurality of features, each feature comprises multiple categories; and 
 a storage storing feature information indicating the categories of the respective features into which the data of samples are classified and class label information indicating whether the samples corresponding to sample objects belong to the specific class, the feature information and the class label information being provided for the samples corresponding to the respective sample objects used for selection of the discriminant feature subset, 
 the feature selector comprising:
 (a) a sample-data extractor extracting part of the samples as training samples and the remaining samples as test samples, from the samples; 
 (b) a classification estimator estimating whether the test samples corresponding to test objects belong to the specific class based on the feature information on the training samples, the class label information on the training samples, and the feature information on the test samples, for respective feature subsets, and determining the correctness of the estimation based on the class label information on the test samples; and 
 (c) a discriminant feature determiner determining the discriminant feature subset based on the correctness of the estimation for each of feature subsets, wherein, 
 
 the classification estimator 
 (b-1) calculates a first conditional probability and a second conditional probability based on the number of the training samples belonging to each category of the respective features, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective features included in the feature subsets for a non-specific class which is a class other than the specific class among classes, and the number of the training samples is counted for each class based on the feature information on the training samples and the class label information on the training samples, 
 (b-2) calculates a first posterior probability that the test sample corresponding to the test object belongs to the specific class based on the feature information on the test sample, a first prior probability of occurrence of the specific class and the first conditional probability, 
 (b-3) calculates a second posterior probability that the test sample corresponding to the test object belongs to the non-specific class based on the feature information on the test sample, a second prior probability of occurrence of the non-specific class and the second conditional probability, 
 (b-4) outputs classification information indicating a result of an estimation on whether the test samples corresponding to the test objects belong to the specific class by comparing the first posterior probability with the second posterior probability, and 
 (b-5) outputs judgement information indicating the correctness of the estimation by comparing the classification information of the test samples with the class label information on the test samples, and 
 the discriminant feature determiner determines the feature subset as the discriminant feature subset, the feature subset corresponds to the judgement information satisfying a predetermined condition among the judgement information of the respective feature subsets. 
 
     
     
       27. A non-transitory computer readable medium encoded with a program for processing information instructing a computer to function as the information processor according to  claim 26 . 
     
     
       28. A method of processing information carried out by an information processor determining whether an input object belongs to a specific class among multiple classes, based on discriminant feature subset consisted of a plurality of discriminant features, each discriminant feature comprises multiple categories, the information processor comprising:
 a storage storing an input pattern corresponding to the input object, a first conditional probability and a second conditional probability, the first conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective discriminant features for the specific class, the second conditional probability is a probability that the data of the input pattern belong to categories corresponding to the respective discriminant features for a non-specific class which is a class other than the specific class among classes; and 
 a classification determiner determining whether the input object belongs to the specific class based on the categories of the respective discriminant features into which data of the input pattern are classified, wherein 
 the input pattern is associated with feature information indicating the categories of the respective discriminant features into which the data of the input pattern are classified, the method comprising: 
 a step of calculating a first posterior probability that the input object belongs to the specific class based on the feature information on the input pattern, a first prior probability of occurrence of the specific class and the first conditional probability; 
 a step of calculating a second posterior probability that the input object belongs to the non-specific class based on the feature information on the input pattern, a second prior probability of occurrence of the non-specific class and the second conditional probability; and 
 a step of determining whether the input object belongs to the specific class by comparing the first posterior probability with the second posterior probability.